Meta's New Mac Coding Agent Costs Up to 20x Less If You Let Meta Train on Your Data

Meta has introduced Muse Code, a terminal coding agent for macOS and Linux, powered by Muse Spark 1.2 – the company's latest model. Meta CEO Mark Zuckerberg highlighted Muse Code's ability to handle extensive software engineering tasks within large repositories, including planning, code generation, and result validation via multiple persistent subagents. This marks Meta's initial major entry into agentic coding, positioning Muse Code as a counterpart to OpenAI's Codex and Anthropic's Claude Code.

Muse Code offers default functionalities like /plan for creating approval-based task plans, /grill for rigorous testing of plans, and /goal for pursuing objectives. It keeps a local event log of all model calls, tool executions, approvals, and edits, enabling it to resume operations post-crash. The coding agent can be installed via a single command in the Mac terminal and does not require a separate app.

Pricing for Muse Code varies based on the level of access granted for training Meta's models. Users who permit Meta to utilize their prompts and completions for future AI model training receive a discounted rate. The Contributor tier, which allows Meta to train its models, costs $0.10 per 1 million input tokens and $0.20 per 1 million output tokens.

This is significantly lower than the Standard tier, priced at $1.25 for input tokens and $4.25 for output tokens, where Meta does not have permission to train its models. The Contributor tier imposes a rate limit of 60 requests per minute and 2,100,000 tokens per minute, while the Standard tier permits 3,000 requests per minute and 4,000,000 tokens per minute.